20240325-中航证券-电子行业周报_英伟达GTC2024召开_关注AI产业链增变量_16页_4mb

NVIDIA Technological Advances in AI and GPU Hardware

Overview

This content discusses NVIDIA's focus on AI-driven technologies, with highlights on the Blackwell GPU architecture, GTC 2024 conference, and performance metrics for AI workloads.

Key Developments

• GTC 2024: Events and announcements centered on AI, featuring the Blackwell GPU and GB200 Superchip.
• Blackwell GPU:
  • Core specifications: 20 peta FLOPS performance in FP4 precision, NVLink support up to 1.8TB/s bandwidth, HBM3e integration with 8TB/s memory bandwidth.
  • Applications: Optimized for large language models (LLMs) and AI computing.
• GB200 Superchip:
  • Architecture: Combines Blackwell GPUs with Grace CPUs (e.g., GB200 NVL72 has 72 GPUs and 36 CPUs).
  • Performance: Achieves up to 1.4 exa FLOPS, enabling scalable AI applications like GPT.
  • Competitor comparison: Outperforms Hopper and Ampere generations in throughput and efficiency.
• HBM and Packaging:
  • HBM3E adoption shows growth with increased bandwidth, supported by 250K/m TSV trend.
  • CoWoS technology facilitates efficient chip integration.
• AI Workloads:
  • Focus on LLMs, requiring real-time processing enhancements.
  • Metrics include energy efficiency targets, such as reduced power per exaFLOP.